[Pass] Refinements of inter-core communication and auto-sync injection logic.

[xiaoyao-NKU]

Improve barrier handling and redundant synchronization elimination logic; Support non-constant coreid in inter-core communication primitives.

[JiaqiGuoSunlune]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Tilelang/src/transform/inject_sunmmio_sync.cc

Lines 1542 to 1544 in 9787b6b

	if (!f->HasNonzeroAttr(tir::attr::kIsGlobalFunc)) {
	return f;
	}

Remove global-func gate that disables sync injection

The new early return on tir.is_global_func causes InjectSunmmioSync to become a no-op in the normal Sunmmio lowering pipeline: OptimizeForTarget invokes this pass at tilelang/engine/phase.py:240, while AnnotateEntryFunc (which sets entry/global function attrs) is only run later at tilelang/engine/phase.py:270. As a result, functions compiled through the default path often lack tir.is_global_func when this pass runs, so wait/barrier insertion is skipped entirely and asynchronous DMA/broadcast ordering is left unsynchronized.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".